Lightweight surface electromyography signal classification method and system

Abstract

The present invention discloses a lightweight surface electromyography signal classification method and system, which includes: obtaining the original surface electromyography signal; maintaining a dynamic reference vector that can track changes in the physiological baseline of the signal; using the dynamic reference vector to standardize the original signal and generate a channel mixing matrix based on the vector; extracting a local time-frequency feature map from the standardized signal; applying the channel mixing matrix to transform the local time-frequency feature map to output a collaborative adaptive feature map that can adapt to changes in muscle coordination patterns; processing the collaborative adaptive feature map through a parallel sequence modeling module that integrates a state-space model and a multi-head self-attention mechanism to obtain the final action category. The present invention improves the robustness of the model to individual differences and muscle fatigue, and has the advantages of lightweight and low latency while maintaining high classification accuracy, and is suitable for mobile or wearable devices.

Description

Technical Field

[0001] The present invention relates to artificial intelligence technology, and in particular to a lightweight surface electromyography signal classification method and system. Background Art

[0002] Swallowing disorder refers to a pathological condition in which food cannot be safely and effectively transported from the mouth to the stomach due to structural or functional damage to related organs such as the mandible, tongue, and throat. This disorder not only seriously affects the patient's quality of life, causing problems such as malnutrition and dehydration, but may also lead to fatal complications such as aspiration pneumonia, placing a heavy medical and economic burden on individuals, families, and even society. Clinical studies have shown that neurodegenerative diseases such as stroke and Parkinson's disease, as well as functional decline associated with aging, are common factors that lead to impaired swallowing function, especially in the elderly population, with a prevalence as high as 15% to 40%. Therefore, the development of an accurate, convenient, and non-invasive swallowing function assessment technology is of vital clinical significance and social value for achieving early diagnosis, guiding rehabilitation treatment, and reducing the risk of complications.

[0003] Currently, there are various assessment methods for dysphagia, each with its own applicable scenarios. Video fluoroscopic swallowing (VFSS), the gold standard for clinical diagnosis, uses X-rays to image the swallowing process in real time, clearly revealing food flow through the throat and aspiration. However, the expensive equipment, radiation exposure, and complex operation make it difficult to adopt as a routine screening tool. Fiberoptic endoscopy (FEES), which involves inserting an endoscope through the nasal cavity, allows for a visual assessment of the anatomy and function of the throat. However, the invasive procedure can cause patient discomfort and requires high operator skill. As a non-invasive alternative, surface electromyography (sEMG) records the electrical activity of swallowing-related muscles to reveal the timing and intensity of muscle activation. Due to its non-invasive, portable, and low-cost nature, it shows great potential for swallowing function assessment. Early sEMG signal analysis relied primarily on manual feature extraction from the time domain, frequency domain, or time-frequency domain, combined with traditional machine learning algorithms (such as support vector machines and K-nearest neighbors) for classification. In recent years, with the development of deep learning technology, convolutional neural networks (CNNs), recurrent neural networks (RNNs), and their variants (such as long short-term memory networks (LSTMs)) have been introduced into sEMG signal processing. These models can automatically learn features and process temporal information, which has improved classification accuracy to a certain extent. Some researchers have also experimented with hybrid models, such as CNN-LSTM or CNN-Transformer, to combine the local feature extraction capabilities of CNNs with the long-term dependency capture capabilities of sequence models, driving technological progress in this field.

[0004] However, the inventors found in their research that although the existing technology has made certain progress, when applying the sEMG signal classification model to real, continuous clinical monitoring or home rehabilitation scenarios, there are still problems such as decreased classification performance due to individual differences and muscle fatigue and high model calculation costs. Summary of the Invention

[0005] The purpose of the invention is to provide a lightweight surface electromyography signal classification method and system in order to solve the above-mentioned problems existing in the prior art.

[0006] The technical solution is a lightweight surface electromyography signal classification method, which is implemented by a preset classification system and includes:

[0007] Acquire multi-channel raw surface electromyographic signals; maintain a dynamic reference vector that can track changes in its physiological baseline;

[0008] The original surface electromyographic signal is normalized by using a dynamic reference vector to generate a normalized electromyographic signal and extract a local time-frequency feature map from it;

[0009] Generate a channel mixing matrix based on the dynamic reference vector;

[0010] Apply the channel mixing matrix to transform the local time-frequency feature map and output the collaborative adaptive feature map;

[0011] Sequence modeling and classification are performed on the collaborative adaptive feature maps to obtain swallowing action categories.

[0012] A lightweight surface electromyography signal classification system, comprising:

[0013] A signal acquisition module is configured to acquire multi-channel raw surface electromyography signals;

[0014] a dynamic baseline maintenance module configured to maintain a dynamic baseline vector capable of tracking changes in a physiological baseline of a raw surface electromyographic signal;

[0015] a normalization module configured to normalize the original surface electromyographic signal using a dynamic reference vector to generate a normalized electromyographic signal;

[0016] A matrix generation module is configured to generate a channel mixing matrix according to a dynamic reference vector;

[0017] a feature extraction module configured to extract a local time-frequency feature map from the normalized electromyographic signal;

[0018] a feature transformation module configured to transform the local time-frequency feature map by applying a channel mixing matrix to output a collaborative adaptive feature map;

[0019] The classification module is configured to perform sequence modeling and classification on the collaborative adaptive feature map to obtain a swallowing action category.

[0020] Beneficial effect: The present invention can improve the robustness of the model to individual differences and muscle fatigue, while maintaining high classification accuracy, reducing the computational cost, and has the advantages of lightweight and low latency. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A flowchart of the steps of a lightweight surface electromyography signal classification method provided by an embodiment of the present invention.

[0022] Figure 2 A flowchart of the steps for maintaining a dynamic reference vector provided by an embodiment of the present invention.

[0023] Figure 3 A flowchart of the steps of generating a channel mixing matrix and outputting a collaborative adaptive feature map provided by an embodiment of the present invention.

[0024] Figure 4 A flowchart of the steps for deriving swallowing action categories provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] To make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the specific embodiments described herein are only for explaining the present invention, rather than for limiting the present invention.

[0026] The research revealed that existing deep learning models commonly suffer from the core technical challenges of weak cross-subject generalization and within-session performance decay. The root of this problem lies in the fact that sEMG signals not only exhibit significant individual heterogeneity due to individual physiological differences (such as muscle morphology and subcutaneous fat thickness), but also undergo dynamic changes due to muscle fatigue during a single use. Muscle fatigue not only causes signal amplitude drift but also triggers compensatory adjustments in the neural center, altering the coordinated activation patterns among multiple muscles. Traditional static normalization methods are unable to track these dynamic changes. When faced with effective compensatory patterns caused by fatigue that have not been seen in the training set, existing models will make progressive misjudgments due to knowledge mismatch, seriously affecting the reliability of the models in practical applications.

[0027] At the model architecture level, efficiently and accurately modeling long-range dependencies in sEMG signals is key to achieving high-accuracy classification, especially when distinguishing subtle swallowing movements. While the Transformer model excels in this area, its quadratic self-attention mechanism leads to high computational costs, making it unsuitable for wearable devices or mobile devices with limited computing power and power consumption. The Mamba model, based on a state-space model (SSM), is designed for efficient processing of long sequences of data. It has demonstrated excellent performance in fields such as speech, audio, and genomics, but its application in sEMG classification and dysphagia assessment remains underexplored. Furthermore, direct application of it to EMG classification tasks (such as dysphagia assessment) also has limitations and yields suboptimal classification results. Specifically, EMG signals often contain high-frequency noise and artifacts (such as those caused by electrode movement or external interference), which can affect the Mamba model's feature extraction capabilities. Designed primarily for long sequences of data in fields such as speech and audio, the Mamba model may not effectively handle high-frequency noise in EMG signals, resulting in reduced classification performance. The assessment of dysphagia requires a model to accurately capture both short-term and long-term temporal dependencies in EMG signals. Although the Mamba model has advantages in modeling long sequences, its selectivity mechanism may not be able to fully adapt to the subtle time-varying features in EMG signals, such as the dynamic changes in muscle activation during swallowing. EMG signals typically come from multiple channels, corresponding to the activities of different muscle groups, and the assessment of dysphagia requires a comprehensive analysis of information from these channels. The Mamba model was originally designed to process sequence data from a single channel or a small number of channels, and lacks an effective fusion mechanism for multi-channel EMG data. Therefore, how to construct a lightweight model that can capture both the state dependencies that Mamba excels at and the global correlations that Transformer excels at without sacrificing computational efficiency is a technical bottleneck that needs to be urgently addressed in this field.

[0028] like Figure 1 As shown in the figure, a lightweight surface electromyography signal classification method is proposed, which is used to classify the user's swallowing movements in real time and accurately. The method is performed by a preset classification system and includes the following steps:

[0029] Acquire multi-channel raw surface EMG signals; maintain a dynamic reference vector that can track changes in its physiological baseline; use the dynamic reference vector to normalize the raw surface EMG signals, generate standardized EMG signals, and extract local time-frequency feature maps from the standardized EMG signals.

[0030] In this embodiment, the raw surface electromyographic signals can be collected in real time by wearing multiple sEMG sensors on the surface of the user's swallowing-related muscles (such as the mandibular and suprahyoid muscles). The raw surface electromyographic signals contain multi-channel time series data of the user's real muscle electrical activity and various noises. In order to cope with the signal differences between different users and the same user in different physiological states, static calibration is not used. Instead, a dynamic reference maintenance module is used to continuously track the drift of the signal baseline and maintain a dynamic reference vector. This vector is used to perform real-time standardization on the raw signal to eliminate amplitude differences and generate standardized electromyographic signals that are easier to process in subsequent modules. The feature extraction module processes the standardized electromyographic signals to extract local time-frequency feature maps that can characterize the details of muscle activity.

[0031] Optionally, preprocessing can also be performed. Filtering techniques (such as bandpass filtering, typically with a frequency range of 20-500 Hz) are applied to remove high-frequency noise and artifacts from the raw surface EMG signals, providing cleaner input data. The raw surface EMG signals, free of high-frequency noise and artifacts, are then segmented into short time segments (e.g., 0.5- to 1-second windows) using a sliding window technique. This technique better captures local features and accommodates the short-term dynamic changes in muscle activity during swallowing.

[0032] According to the dynamic reference vector, a channel mixing matrix is ​​generated; the channel mixing matrix is ​​applied to transform the local time-frequency feature map and output a collaborative adaptive feature map.

[0033] In this embodiment, this step is used to achieve adaptation of the muscle fatigue compensation mode. Since the dynamic reference vector not only reflects the signal amplitude baseline, but also contains information about the user's overall physiological fatigue state. Therefore, the matrix generation module takes the dynamic reference vector as input to generate a channel mixing matrix. The role of the channel mixing matrix is ​​to act as a dynamic adapter or translator. The feature transformation module applies the matrix to perform a linear transformation on the local time-frequency feature map obtained in the previous step. This process can dynamically adjust the relationship between different muscle channel features according to the user's current fatigue state, so that the effective compensation mode caused by fatigue, which the model was originally unfamiliar with, is reconstructed into a mode that the model can understand, thereby generating a collaborative adaptive feature map.

[0034] Sequence modeling and classification are performed on the collaborative adaptive feature maps to obtain swallowing action categories.

[0035] In this embodiment, this step is used to perform deep temporal modeling and final decision-making on the features after front-end adaptive processing. The collaborative adaptive feature graph is input into the classification module, which first uses the sequence modeling unit to capture the complex long-range temporal dependencies in the feature graph. The classification head unit then maps the sequence features to one of multiple preset swallowing action categories, such as normal swallowing, effortful swallowing, and swallowing with aspiration risk, completing the entire classification task.

[0036] Through the above steps, this embodiment can achieve dynamic calibration, feature adaptive adjustment and high-precision classification of sEMG signals within a unified framework, and is particularly suitable for real and changing clinical or home application scenarios.

[0037] like Figure 2 As shown, according to one aspect of the present application, in order to ensure that the dynamic reference vector can accurately and stably track changes in the physiological baseline, the maintenance process of the dynamic reference vector specifically includes:

[0038] Determine whether the original surface electromyographic signal corresponds to the resting state of the muscle; when it is determined to correspond to the resting state of the muscle, use the original surface electromyographic signal to update the dynamic reference vector. The specific steps are as follows: from the fragment of the original surface electromyographic signal in the resting state of the muscle, calculate the energy characteristics that represent the current energy level of the fragment; through the exponential moving average algorithm, weightedly fuse the energy characteristics with the dynamic reference vector at the previous moment to achieve smooth update of the dynamic reference vector.

[0039] In this embodiment, a gating mechanism is required to determine the signal state. Optionally, the mechanism can be implemented through an energy threshold: when the energy of the signal in a small time window is lower than a preset threshold, it is determined to be a muscle resting state. In addition, methods such as first-order difference can also be used to detect the stationarity of the signal. When it is determined to be a resting state, the update process is triggered. It is necessary to calculate the energy feature to quantify the baseline level of the current resting signal. Preferably, the energy feature can be the root mean square (RMS) value of the signal, which can robustly reflect the average power of the signal. The exponential moving average (EMA) algorithm is used for updating. The calculation formula is: DB t =α·E t +(1-α)·DB t-1 ; Among them DB t and DB t-1 are the dynamic reference vectors at the current and previous moments, E tis the currently calculated energy signature, α is the smoothing factor, and the value range of α is (0, 1). Preferably, α is a smaller value, such as between 0.01 and 0.05. A smaller α value means that the update process relies more on the historical benchmark, making the changes in the dynamic benchmark vector very smooth, effectively filtering out the interference of transient noise, and only tracking slow, long-term baseline drift caused by factors such as fatigue. Through this conditional, smooth update mechanism, the accuracy and robustness of the dynamic benchmark vector are improved, providing high-quality input for subsequent standardization and adaptation modules.

[0040] In another embodiment of the present application, the energy feature may also be the mean absolute value (MAV), integrated electromyographic value (iEMG), waveform length (WL) or zero crossings (ZC), or other statistics that are well known in the art and can reflect the signal energy or complexity.

[0041] In another embodiment of the present application, the mechanism for determining the resting state of the muscle may also be a lightweight, pre-trained binary classification model, such as a logistic regression classifier or an extremely small neural network, to achieve more accurate state determination.

[0042] like Figure 3 As shown, according to one aspect of the present application, the process of generating a collaborative adaptive feature map specifically includes:

[0043] The dynamic reference vector is parsed by a parameter generation network to construct a channel mixing matrix, which is used to encode the muscle synergy pattern indicated by the physiological baseline changes. The local time-frequency feature map is linearly reconstructed through the channel mixing matrix to adaptively adjust the dependencies between the feature channels and obtain a collaborative adaptive feature map.

[0044] In this embodiment, the core of the process lies in establishing a dynamic mapping from physiological state to feature adjustment strategies. Specifically, the parameter generation network is preferably a lightweight multilayer perceptron (MLP). For example, it can be a simple network consisting of two fully connected layers. This MLP does not process high-dimensional electromyographic features, but instead accepts as its sole input the low-dimensional dynamic reference vector generated in the previous step. Because the dynamic reference vector itself is a quantitative representation of the system's current fatigue state, the MLP's task is to learn a nonlinear mapping function from fatigue state to changes in inter-channel synergy. The output of the MLP is the channel mixing matrix. Mathematically, if there are N sEMG channels, this matrix is ​​an N×N matrix. Through matrix multiplication, this channel mixing matrix is ​​applied to linearly reconstruct the local time-frequency feature maps of the N channels. In essence, this dynamically weights and combines features from different muscle channels. For example, when the system detects fatigue in a core muscle group (as indicated by changes in the dynamic reference vector), the generated channel mixing matrix may automatically reduce the feature weight of the fatigued channel while increasing the feature weights of other channels that have a compensatory effect. This enables the model to understand fatigue and proactively adapt, rather than passively treating fatigue compensation patterns as errors. This improves the model's stability and reliability in long-term, continuous use scenarios.

[0045] like Figure 4 As shown, according to one aspect of the present application, the sequence modeling and classification steps are further refined, and the steps specifically include:

[0046] The collaborative adaptive feature map is fed into the state-space model unit and the multi-head self-attention unit in parallel; the state-space model unit captures the temporal causal dependency in the collaborative adaptive feature map to generate state-space features; the multi-head self-attention unit captures the global position association within the collaborative adaptive feature map to generate global attention features; the state-space features and the global attention features are fused to form a fused sequence feature; the classification head module determines the swallowing action category based on the fused sequence features.

[0047] In this embodiment, to address the conflict between efficiency and performance in existing technologies, a parallel fusion architecture is adopted within the classification module. Specifically, the collaborative adaptive feature map is simultaneously fed into two parallel processing units. The first is a state-space model unit, preferably employing the Mamba model or its variants. This unit excels at processing long sequences with linear time complexity. Through its internal recurrent state mechanism, it can efficiently capture the causal relationships and state changes of signals over time, such as the complete dynamic process of muscle activation, force generation, and relaxation. Its output is state-space features. The second is a multi-head self-attention unit, preferably employing the standard multi-head self-attention module in the Transformer encoder. For example, four attention heads can be set, with a dropout rate of 0.2. This unit can overcome the limitations of time series and calculate the correlation between any two time points in the feature map, thereby accurately capturing the global collaborative patterns that are critical for distinguishing fine movements. Its output is a global attention feature. The feature fusion unit combines the state-space features with the global attention features. Optionally, the fusion method can be element-wise addition, concatenation or weighted summation. Preferably, this embodiment adopts element-wise addition to obtain information-complementary fusion sequence features without significantly increasing the number of parameters. The fusion sequence features are sent to the classification head module, which is usually composed of a global average pooling layer and one or more fully connected layers to finally determine the swallowing action category. Through this parallel and complementary design, while maintaining low computing latency and lightweight model, it takes into account the ability to capture deep features of two different properties (causal temporal dependency and non-causal global correlation), thereby improving the accuracy of classification.

[0048] In another embodiment of the present application, the feature fusion unit includes an attention fusion module. Multiple channels of the neck electromyography (EMG) signal correspond to different muscle groups (such as the mandibular muscles and hyoid muscles), some of which play a dominant role in the swallowing process. Therefore, an attention-based fusion module, the Multi-Channel Spatial Attention Selection (MCSAS) module, was constructed. By learning the correlations between different channels, it dynamically adjusts the contribution weights of each channel to more accurately reflect the synergistic effects of muscle groups during swallowing. Integrating the MCSAS channel attention module into the Mamba model enhances the model's ability to focus on key time periods in the EMG signal, strengthens the model's ability to process multi-channel data, and further improves its EMG signal classification performance. Specifically, based on swallowing physiology research, channel weights for key muscle groups are predefined to serve as initialization or regularization constraints for the attention mechanism. Spatial correlations between channels are modeled through convolution to identify muscle group channels that are highly correlated with swallowing movements. Dynamically assigning weights to signal segments related to swallowing movements improves the model's ability to model temporal dependencies.

[0049] In another embodiment of the present application, the feature fusion unit may also adopt a method of feature concatenation followed by 1x1 convolution for dimensionality reduction, or adopt a learnable gating unit to dynamically adjust the fusion weight of the state space feature and the global attention feature.

[0050] According to one aspect of the present application, the process of extracting local time-frequency feature maps is implemented by an integrated multi-scale attention residual module (MARBlock), specifically including:

[0051] The standardized EMG signal is processed by a deformable convolution unit, and the temporal features are captured in parallel using convolution kernels of various sizes and fused into multi-scale fusion features. The multi-scale fusion features are analyzed by a channel attention unit to generate a set of channel weights, which are then used to modulate the multi-scale fusion features to highlight key channel information and obtain channel-weighted features. The channel-weighted features are combined with the standardized EMG signal through a residual connection to generate a local time-frequency feature map.

[0052] Optionally, the deformable convolution unit includes:

[0053] Predetermining parallel convolution paths, wherein each convolution path is configured with a convolution kernel of a different size, for extracting features of different time scales from the normalized electromyographic signal;

[0054] The splicing processing unit is used to aggregate the features output by a predetermined number of parallel convolution paths along the channel dimension to generate multi-scale fusion features.

[0055] Optionally, the channel attention unit includes:

[0056] Compression unit, used to perform global information compression on multi-scale fusion features and extract channel descriptors;

[0057] An excitation unit, configured to perform a nonlinear transformation on the channel descriptor through at least one fully connected layer to learn the channel weights;

[0058] The recalibration unit is used to apply channel weights to the multi-scale fusion features channel by channel to generate channel-weighted features.

[0059] In this embodiment, the deformable convolution unit is designed to include multiple parallel convolution paths. For example, three parallel 1D convolution paths can be set up, with kernel sizes of 3, 5, and 7, respectively. This enables the unit to simultaneously capture patterns at different timescales in the sEMG signal, such as short muscle activation spikes (captured by small kernels) and longer rhythmic contractions (captured by large kernels). The output feature maps of each path are then aggregated along the channel dimension in a concatenation processing unit to form multi-scale fused features containing rich scale information. The multi-scale fused features are then fed into a channel attention unit. This unit preferably adopts a squeeze-and-excitation (SE) architecture, comprising a compression unit, an excitation unit, and a recalibration unit. Specifically, the compression unit performs global average pooling (squeeze) on the input features, compressing the temporal information of each channel into a single value to obtain a channel descriptor. The excitation unit (typically consisting of two fully connected layers and a nonlinear activation function) processes this channel descriptor to learn the importance weights (excitation) for each channel. The recalibration unit multiplies these weights (after Sigmoid activation) channel by channel back to the original multi-scale fusion features, thereby modulating the features, enhancing the expression of useful channels, and suppressing noise or redundant channels. The modulated channel weighted features are added element-by-element to the original input of the module (normalized electromyographic signals, or signals after dimension matching through 1x1 convolution) through residual connection (skip connection) to generate a local time-frequency feature map. This helps to alleviate the gradient vanishing problem in deep networks and increase the trainability of the network. Through the above-mentioned integrated design, this embodiment can efficiently extract local time-frequency features with multi-scale and high discriminability in a single forward propagation, providing high-quality input for subsequent classification tasks.

[0060] Optionally, the process of generating multi-scale fusion features can also be: creating a deformable multi-scale parallel (InceptDeform) module. The traditional multi-scale parallel convolution (Inception) module contains multiple parallel branches, such as 1x1 convolution, 3x3 convolution, 5x5 convolution and pooling layer. Each branch is responsible for extracting features of different scales. Although it can extract multi-scale features, its fixed receptive field is difficult to fully adapt to the dynamic changes in these modal signals. Therefore, deformable convolution is introduced into the traditional Inception module, and the convolution operation of each branch is changed to deformable convolution, so that the receptive field of the convolution kernel can dynamically adjust the shape and position according to the input data. This module uses a multi-branch parallel architecture to process the input feature map (dimension: batch size B × number of input channels C in × time length T), through four parallel operations: 1×1 deformable convolution: generates spatial offsets (Offset) to capture local subtle features; 3×3 deformable convolution: generates offsets with larger receptive fields to extract medium-range features; 5×5 decomposition convolution: first performs 5×1 vertical convolution, then 1×5 horizontal convolution, reducing the number of parameters while expanding the receptive field; 3×3 maximum pooling: extracts the most significant feature responses and enhances feature robustness. All branch outputs are concatenated (Concat) in the channel dimension, and the number of output channels is expanded to 4 times the input (4C out ), generating multi-scale fused features while retaining the original time dimension T. This approach is suitable for the Mamba model for time-sensitive tasks (such as video analysis and speech processing). This embodiment retains the multi-scale feature extraction capabilities of the Inception module while increasing adaptability. The model can adaptively focus on key areas of each modality signal, more accurately capturing complex patterns associated with swallowing movements. This reduces reliance on large-scale annotated data and improves the robustness of feature extraction.

[0061] In another embodiment of the present application, a lightweight model is proposed based on a lightweight surface electromyography signal classification method. The overall architecture of the model adopts a hierarchical coding structure, including a multi-scale attention residual (MAR) module and a pooling and classification layer. The multi-scale attention residual module (MARBlock) integrates multi-scale temporal convolution, channel attention mechanism and regularized residual learning into a unified module. Multi-scale temporal convolution is to extract multi-scale temporal features. Its core component is a parallel multi-scale convolution (MultiScaleConv) layer, which uses a deformable convolution kernel to synchronously capture short-term transient patterns (such as muscle activation spikes) and long-term rhythmic trends (such as breathing cycles). For the splicing and fusion of input feature maps into rich representations, F multi = Concat(Conv1D (k-3) (F), Conv1D (k-5)(F), Conv1D (k-7) (F)); where F multi is the feature map after multi-scale fusion, Concat is the splicing function, Conv1D (k-3) The convolution operation is performed with a one-dimensional convolution kernel size of 3, and F is the original input feature map. The dynamic channel attention mechanism is inspired by SENet. The module embeds a compression excitation module (SEBlock) after the deformable convolution to adaptively recalibrate the channel response. Input feature F multi Statistics are aggregated through global temporal pooling (GAP), followed by two learnable fully connected layers with a dimension ratio r: s = σ(W2·δ(W1·GAP (F multi ), where W1 and W2 are the fully connected weight matrices of the first and second layers, respectively; σ is the activation function, typically Sigmoid; δ is the activation function, typically ReLU; and s is the channel-wise attention weight vector. The generated attention weights are used to enhance discriminative channels and suppress noise. Regularized residual learning introduces group normalization (GroupNorm) and dropout to ensure training stability and generalization. Group normalization ensures robustness to small batches of data, while dropout (applied before residual addition) prevents excessive coupling of redundant features. The final output fuses the processed features with the original input via skip connections. Over three layers, the feature resolution is sequentially increased from T to T / 2 to T / 4 to T / 8, and the number of channels from hidden to hidden to hidden2 to hidden4, extracting richer local and mid-range patterns layer by layer while controlling the computational effort. Where T is the temporal length of the input signal, and hidden is the number of channels per layer.

[0062] The pooling and classification layer uses adaptive average pooling (AdaptiveAvgPool) to globally average the encoded time series features in the time dimension, automatically compressing sequences of any length into the shape [B, hidden dim , 1] tensor; flatten it into [B, hidden dim ] vector, and then pass through the fully connected layer Linear(hidden dim , num classes ) maps these global features to the scores of each category, completing the transformation from time series features to the final classification results. Where B is the batch size, hidden dim is the feature dimension after encoding, num classes is the number of categories.

[0063] According to one aspect of the present application, a lightweight surface electromyography signal classification system capable of executing the above-mentioned embodiment is also provided. The system includes: a signal acquisition module, a dynamic reference maintenance module, a normalization module, a matrix generation module, a feature extraction module, a feature transformation module, and a classification module.

[0064] In this embodiment, the signal acquisition module is used to acquire multi-channel original surface electromyography signals;

[0065] A dynamic reference maintenance module, preferably having a gated dynamic update mechanism implemented therein, for maintaining a dynamic reference vector capable of tracking changes in the physiological baseline of the original surface electromyographic signal;

[0066] The normalization module is used to normalize the original signal using a dynamic reference vector to generate a normalized electromyographic signal;

[0067] The matrix generation module works in conjunction with the feature transformation module. Specifically, the matrix generation module preferably includes a multi-layer perceptron as a parameter generation network, which is used to generate a channel mixing matrix based on the dynamic reference vector; the feature transformation module uses this matrix to linearly reconstruct the features;

[0068] A feature extraction module, preferably implemented internally as one or more multi-scale attention residual (MAR) modules, for extracting local time-frequency feature maps from the normalized EMG signal;

[0069] The classification module preferably implements a parallel fusion architecture. Specifically, the module can be further divided into a state-space model unit, a multi-head self-attention unit, a feature fusion unit, and a classification head unit to achieve efficient and accurate sequence modeling and classification to derive swallowing action categories.

[0070] In another embodiment of the present application, the matrix generation module includes a parameter generation network for parsing the dynamic reference vector to construct a channel mixing matrix;

[0071] The feature transformation module is configured to transform (linearly reconstruct) the local time-frequency feature map through a channel mixing matrix to obtain a collaborative adaptive feature map.

[0072] In another embodiment of the present application, the classification module includes:

[0073] The state-space model unit and the multi-head self-attention unit are configured to receive the collaborative adaptive feature map in parallel and generate state-space features and global attention features respectively;

[0074] A feature fusion unit for fusing state space features with global attention features to form fused sequence features; and

[0075] The classification head unit is used to determine the swallowing action category based on the fusion sequence features.

[0076] In another embodiment of the present application, the feature extraction module includes:

[0077] Deformable convolution unit, used to process normalized EMG signals to generate multi-scale fusion features;

[0078] Channel attention unit, used to modulate multi-scale fusion features to generate channel-weighted features;

[0079] The residual fusion unit is used to combine the channel weighted features with the normalized EMG signal to generate a local time-frequency feature map.

[0080] The system can be implemented as a dedicated hardware device, for example, integrated into a portable medical monitor or smart wearable device (e.g., a smart neck brace or armband). Furthermore, the system's functional modules can also be implemented as software instructions, stored on a computer-readable storage medium, and loaded and executed by one or more processors to implement the methods described in the present invention on a general-purpose computing device (e.g., a smartphone, tablet, or personal computer).

[0081] In some embodiments, the dynamic reference maintenance module is specifically implemented as a gated dynamic reference calibration (GDBC) module, including a gating unit and a reference update unit; the matrix generation module and the feature transformation module jointly constitute a fatigue-aware collaborative adaptive module (FASAM).

[0082] In some embodiments, the classification system adopts a hierarchical architecture, including: a signal preprocessing layer (GDBC module); a feature extraction layer (MAR module); an adaptive adjustment layer (FASAM module); a sequence modeling layer (parallel Mamba-MHA architecture); and a classification decision layer.

[0083] In a preferred embodiment of the present invention, a dynamic baseline maintenance module, a normalization module, a matrix generation module, and a feature transformation module together constitute a front-end adaptive system. This system performs in-depth, personalized calibration and reconstruction of input signals before the features are fed into the core sequence modeling and classification modules. Specifically, the dynamic baseline maintenance module is internally implemented as a gated dynamic baseline calibration (GDBC) module, primarily composed of a gating unit and a baseline update unit. The gating unit is configured to determine in real time whether the input raw surface electromyographic signal is in a muscle task phase or a resting phase. The baseline update unit is activated only when the gating unit determines that the input is in a resting phase. It calculates an energy signature (such as the RMS value) representing the current energy level from the raw signal segments in the resting state. This energy signature is then weighted and fused with a previously stored dynamic baseline vector using an exponential moving average algorithm to smoothly update the dynamic baseline vector. The continuously maintained dynamic baseline vector is used by the normalization module for real-time normalization of the raw signal and, as a key physiological status indicator, is fed into subsequent modules.

[0084] Furthermore, the matrix generation module and the feature transformation module work together to form the Fatigue-Aware Synergistic Adaptive Module (FASAM). The matrix generation module contains a lightweight multi-layer perceptron (MLP) as a parameter generation network. This network uniquely takes a dynamic reference vector as input, parses the overall fatigue state information contained therein, and constructs a channel mixing matrix. The feature transformation module applies this real-time generated channel mixing matrix to perform a linear transformation on the local time-frequency feature map output by the upstream feature extraction module. This process essentially dynamically reorganizes and weights the features of different muscle channels based on the user's current fatigue level, ultimately outputting a collaborative adaptive feature map that can adapt to changes in muscle coordination patterns for processing by the final classifier.

[0085] Case 1: A wearable neck electromyography acquisition device is used to conduct a real-time assessment of the swallowing function of a post-stroke rehabilitation patient during a water drinking test. Specifically, a rehabilitation therapist guides a post-stroke patient to perform a quantitative water drinking test. The patient wears a flexible neck ring that integrates four sEMG sensing channels. The sensors are attached to the mylohyoid muscle and sternohyoid muscle areas on both sides. The flexible neck ring transmits the collected sEMG data wirelessly to a tablet computer via low-power Bluetooth. The classification system software of this application is run on the tablet computer. The sEMG signal acquisition frequency is 1000Hz. The system processes the data with a sliding window of 1 second (1000 sampling points). Therefore, at any moment, the original surface electromyography signal X input to the system is a tensor of size [1, 4, 1000]. When working, the process is as follows:

[0086] Step 1: Before the patient starts the water drinking test, the system has been running for a while in a static state and initialized the dynamic reference vector DB. Assume that at time t-1, the vector value is DB t-1 =[0.15, 0.12, 0.16, 0.13]. At time t, the new signal window X t The system first calculates X t The overall energy of (e.g., the average of the RMS values ​​of all channels) is 0.04, which is lower than the preset resting threshold of 0.05, so the window is judged to be in the resting state of the muscle. t The energy characteristics E of each channel t , preferably the RMS value, to obtain E t =[0.04, 0.035, 0.042, 0.038]. The exponential moving average algorithm is used to update the dynamic reference vector. The smoothing factor α is set to 0.01. The update process is as follows: DB t =α·E t +(1-α)·DB t-1 ;DB t =0.01·[0.04, 0.035, 0.042, 0.038]+0.99·[0.15, 0.12, 0.16, 0.13]; DB t =[0.0004, 0.00035, 0.00042, 0.00038]+[0.1485, 0.1188, 0.1584, 0.1287]; DB t =[0.1489, 0.11915, 0.15882, 0.12908]; newly calculated DB t It will be stored by the system and used for standardization of subsequent task status signals.

[0087] Step 2: Assume that at time t+k, the patient performs a swallowing action and the system collects the signal window X of the task state t+k The signal is represented by the current dynamic reference vector DB t Normalization is performed to obtain a standardized EMG signal. The standardized EMG signal is fed into one or more MAR modules. Within a MAR module: the input features (assuming the size is [1, 64, 500]) are fed into three 1D convolutional layers in parallel, with convolution kernel sizes of 3, 5, and 7, respectively. The output features of the three convolutional layers (all [1, 64, 500]) are concatenated along the channel dimension to form a multi-scale fusion feature F with a size of [1, 192, 500]. ms . Multi-scale fusion feature F ms Perform global average pooling GAP(F ms), compressing it from [1, 192, 500] to a channel descriptor z of size [1, 192]. The channel descriptor z is processed by a two-layer fully connected network. The first layer W1 compresses the dimension from 192 to 48 (compression ratio r = 4) and is activated by the ReLU function δ; the second layer W2 restores the dimension from 48 to 192 and is activated by the Sigmoid function σ. The calculation process is s = σ(W2·δ(W1·z)), and the channel attention weight s of size [1, 192] is obtained. The weight s is combined with F ms Multiply channel by channel to obtain the channel weighted feature F cw =s*F ms . Channel weighted feature F cw Added to the input of the MAR module (after dimension matching), the final output is a local time-frequency feature map F with a size of [1, 64, 500] ltf .

[0088] Step 3: Local time-frequency feature map F ltf While being sent downstream, the system uses the current dynamic reference vector DB t (size is [1, 4]) to generate the adjustment strategy. Dynamic benchmark vector DB t It is input to the multi-layer perceptron MLP as the parameter generation network. The MLP contains two hidden layers (for example, 4->16->16), and its output is reshaped into a 4x4 channel mixing matrix CMM t The calculation process is CMM t =MLP(DB t ). To simplify the explanation, assume that the local time-frequency feature map F ltf The number of channels is also 4. Apply the channel mixing matrix CMM t For the local time-frequency feature map F ltf Perform linear transformation (matrix multiplication) and coordinate adaptive feature map F sa =CMM t ·F ltf This process reconstructs the features of the four channels and outputs the collaborative adaptive feature map F sa .

[0089] Step 4: Collaborative Adaptive Feature Map F sa It is input to the state space model unit Mamba and the multi-head self-attention unit MHA at the same time. The Mamba unit processes the collaborative adaptive feature map F sa , output state space feature F mamba The MHA unit processes the collaborative adaptive feature map F sa , output global attention feature F mha The two parallel output features are added element by element to obtain the fused sequence feature F fusedThe calculation process is F fused =F mamba +F mha . Fusion sequence feature F fused After global pooling and the final fully connected classification head, the output represents the probability distribution of each swallowing action category, for example: {normal swallowing: 0.1, effortful swallowing: 0.85, delayed swallowing: 0.05}.

[0090] Through the complete process described in this embodiment, the system can accurately assess whether a swallowing attempt is an effortful swallow based on the patient's real-time sEMG signals. More importantly, even when the patient experiences muscle fatigue due to repeated testing, resulting in changes in the amplitude and coordination pattern of the EMG signals, the system can maintain the high stability and accuracy of the assessment results through the dynamic adaptive mechanisms of GDBC and FASAM, providing a reliable quantitative basis for clinical diagnosis and rehabilitation treatment.

[0091] Case 2: The effectiveness of this application was verified using the PyTorch 1.13 deep learning framework on a server equipped with an NVIDIA RTX 4090 GPU and 128GB of memory. The dataset used was a publicly available sEMG swallowing dataset, or an internal dataset if available. Performance evaluation metrics included accuracy, F1-score, model parameters, and inference latency. The loss function used was the cross-entropy loss function. The optimizer used was the AdamW optimizer with weight decay. For example, the learning rate was set to 1e-4 and the weight decay was set to 0.01. For example, the batch size was set to 64 and the total number of training epochs was 100. The learning rate scheduling strategy used a cosine annealing learning rate scheduler to fine-tune the model in the later stages of training. To improve the robustness of the model, data augmentation was performed on the sEMG signals before training, such as adding a small amount of Gaussian white noise or performing random time scaling. A comparative CNN-Transformer hybrid model was set up. This model also uses a CNN to extract local features and a Transformer to capture long-term dependencies. The experiments were as follows: Model A: CNN-Mamba model; Model B: Based on Model A, with the addition of a MAR module; Model C (the complete solution of this application): Based on Model B, with the addition of a GDBC+FASAM adaptive mechanism. The experimental results are as follows: When the model is CNN-Transformer (comparison), the accuracy is 92.1%, the F1 score is 91.8%, the parameter amount is 16.59MB, and the delay is 51.3ms; when it is model A, the accuracy is 93.5%, the F1 score is 93.2%, the parameter amount is 5.10MB, and the delay is 24.5ms; when it is model B, the accuracy is 94.6%, the F1 score is 94.3%, the parameter amount is 5.42MB, and the delay is 25.1ms; when it is model C, the accuracy is 95.8%, the F1 score is 95.5%, the parameter amount is 5.85MB, and the delay is 25.8ms. Compared with the comparison example, this application has improved both the accuracy and F1 score, while reducing the model parameter amount and inference delay. Experiments have shown that the introduction of each step of the MAR module and the GDBC+FASAM module has brought about a steady improvement in performance.

[0092] In summary, to address the challenges of weak cross-subject generalization and within-session performance degradation in real-world scenarios, a two-layer adaptive mechanism consisting of gated dynamic baseline calibration (GDBC) and a fatigue-aware collaborative adaptation module (FASAM) is proposed to achieve deep adaptation to individual differences and dynamic muscle fatigue. The GDBC module addresses signal amplitude drift caused by fatigue. Unlike traditional static, one-time calibration, this module performs real-time state determination on the raw surface EMG input, distinguishing between muscle activity and resting periods. It uses an exponential moving average algorithm to smoothly update a dynamic baseline vector only during signal segments identified as resting periods. This vector thus serves as a dynamic anchor for tracking the user's individual physiological baseline and fatigue status. Throughout the session, all raw signals are normalized using this real-time, updated dynamic baseline vector, effectively compensating for amplitude degradation caused by fatigue and ensuring consistency in the energy scale of the signal input to subsequent models.

[0093] On this basis, to address the deeper issue of changes in muscle coordination patterns caused by fatigue compensation, the FASAM module was further introduced. This module receives as input the dynamic reference vector generated in the previous stage that characterizes the overall fatigue state of the system, and a lightweight parameter generation network parses it into a channel mixing matrix. The matrix is ​​then applied to linearly transform the local time-frequency feature maps extracted by the deep network, dynamically reconstructing and adjusting the relationship between different muscle channel features based on the current level of fatigue. The effective compensation pattern caused by fatigue, which the model originally did not recognize, is translated into a coordination pattern that is closer to the coordination pattern learned in a healthy state, allowing the classifier to correctly understand and identify it, thus resolving the problem of progressive misjudgment caused by fatigue.

[0094] When modeling long sequences of data such as surface electromyography (EMG), existing methods struggle to balance computational efficiency and global model expressiveness. Therefore, a parallel fusion sequence modeling architecture is designed to leverage the complementary strengths of a state-space model (Mamba) and multi-head self-attention (MHA). Replacing the computationally expensive Transformer model with the computationally efficient Mamba model may compromise the ability to capture the critical, non-causal, global muscle co-activations in swallowing. To this end, this approach utilizes multi-scale feature extraction, adaptive noise suppression, temporal modeling and attention mechanisms, multi-channel fusion, and interpretable feature extraction to overcome the complexities and challenges of cervical EMG signal analysis. This improves the accuracy and robustness of the feature extraction module, providing a more reliable and practical tool for the assessment and treatment of dysphagia. The state-space model unit, leveraging its efficient recurrent state mechanism, excels at capturing causal dependencies in the temporal evolution of signals, such as the complete process from muscle activation to relaxation. Meanwhile, the multi-head self-attention unit, unconstrained by temporal order, focuses on computing the correlation strength between any two time points in the feature map, accurately capturing global, position-independent muscle co-activation patterns. The state-space features representing temporal causal dependencies and the global attention features representing global positional associations are effectively fused to form fused sequence features with higher information density and richer expression dimensions, which are then processed by the final classification head. This parallel design enables the model to benefit from the linear time complexity and low memory consumption of Mamba while retaining the powerful global information integration capabilities of the attention mechanism. This achieves an optimal balance between classification performance and computational efficiency, making it suitable for resource-constrained mobile and wearable applications.

[0095] The preferred embodiments of the present invention are described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the scope of protection of the present invention.

Claims

1. A lightweight surface electromyography signal classification method, performed by a preset classification system, characterized in that: include: Acquire multi-channel raw surface electromyography signals; Maintain a dynamic reference vector that can track changes in its physiological baseline; The original surface electromyographic signal is normalized by using a dynamic reference vector to generate a normalized electromyographic signal and extract a local time-frequency feature map from it; Generate a channel mixing matrix based on the dynamic reference vector; Apply the channel mixing matrix to transform the local time-frequency feature map and output the collaborative adaptive feature map; Perform sequence modeling and classification on the collaborative adaptive feature map to obtain swallowing action categories; Generate channel mixing matrix and output co-adaptive feature map, including: The dynamic reference vector is parsed by the parameter generation network to construct the channel mixing matrix; The local time-frequency feature map is linearly reconstructed through the channel mixing matrix, and the dependency between feature channels is adaptively adjusted to obtain a collaborative adaptive feature map. The parameter generation network is a multi-layer perceptron; the multi-layer perceptron receives a dynamic reference vector as input, and its output is the channel mixing matrix.

2. The method according to claim 1, characterized in that Maintain a dynamic reference vector, including: The gating mechanism is used to determine whether the raw surface electromyographic signal corresponds to the resting state of the muscle; If so, the dynamic reference vector is updated using the original surface electromyography signal.

3. The method according to claim 1, characterized in that Derive swallowing action categories, including: Feed the co-adaptive feature map into the state-space model unit and the multi-head self-attention unit in parallel; The state space model unit captures the temporal causal dependency in the collaborative adaptive feature graph to generate state space features; The multi-head self-attention unit captures the global position correlation in the collaborative adaptive feature map and generates the global attention feature; Fuse the state space features with the global attention features to form fused sequence features; The classification head module determines the swallowing action category based on the fused sequence features.

4. The method according to claim 1, wherein Extract local time-frequency feature maps, including: The normalized EMG signal is processed by a deformable convolution unit, and convolution kernels of different sizes are used to capture temporal features in parallel and fuse them into multi-scale fusion features. The channel attention unit analyzes the multi-scale fusion features, generates channel weights, and uses them to modulate the multi-scale fusion features to obtain channel weighted features; The channel weighted features are combined with the normalized electromyographic signals through residual connections to generate local time-frequency feature maps.

5. The method according to claim 2, characterized in that The dynamic reference vector is updated using the raw surface electromyography signal, including: From the segments of raw surface electromyographic signals at resting state of the muscle, energy signatures representing its current energy level are calculated; The energy feature is weightedly fused with the dynamic reference vector of the previous moment through the exponential moving average algorithm to achieve smooth update of the dynamic reference vector; the update formula is: DB t =α·E t +(1-α)·DB t-1 ; Among them DB t and DB t-1 are the dynamic reference vectors at the current and previous moments, E t is the currently calculated energy feature, and α is the smoothing factor.

6. The method according to claim 4, characterized in that The deformable convolution unit includes: Predetermining parallel convolution paths, wherein each convolution path is configured with a convolution kernel of a different size, for extracting features of different time scales from the normalized electromyographic signal; The splicing processing unit is used to aggregate the features output by a predetermined number of parallel convolution paths along the channel dimension to generate multi-scale fusion features.

7. The method according to claim 4, characterized in that The channel attention unit includes: Compression unit, used to perform global information compression on multi-scale fusion features and extract channel descriptors; An excitation unit, configured to perform a nonlinear transformation on the channel descriptor through at least one fully connected layer to learn the channel weights; The recalibration unit is used to apply channel weights to the multi-scale fusion features channel by channel to generate channel-weighted features.

8. A lightweight surface electromyography signal classification system, characterized in that: include: A signal acquisition module is configured to acquire multi-channel raw surface electromyography signals; a dynamic baseline maintenance module configured to maintain a dynamic baseline vector capable of tracking changes in a physiological baseline of a raw surface electromyographic signal; a normalization module configured to normalize the original surface electromyographic signal using a dynamic reference vector to generate a normalized electromyographic signal; A matrix generation module is configured to generate a channel mixing matrix according to a dynamic reference vector; Generating the channel mixing matrix includes: parsing the dynamic reference vector by a parameter generation network to construct the channel mixing matrix; the parameter generation network is a multi-layer perceptron; the multi-layer perceptron receives the dynamic reference vector as input, and its output is the channel mixing matrix; a feature extraction module configured to extract a local time-frequency feature map from the normalized electromyographic signal; The feature transformation module is configured to transform the local time-frequency feature map by applying a channel mixing matrix to output a collaborative adaptive feature map; wherein outputting the collaborative adaptive feature map includes: linearly reconstructing the local time-frequency feature map by using the channel mixing matrix, adaptively adjusting the dependency between feature channels, and obtaining the collaborative adaptive feature map; The classification module is configured to perform sequence modeling and classification on the collaborative adaptive feature map to obtain a swallowing action category.

Images